3D convolutional neural network based on spatial-spectral feature pictures learning for decoding motor imagery EEG...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Non-invasive brain-computer interfaces (BCI) show potential in neurorehabilitation.
• Motor imagery electroencephalography (EEG) signals suffer from low signal-to-noise ratio and limited resolution.
• Existing deep neural networks often neglect combined spatial-frequency EEG features, limiting motor imagery decoding accuracy.

Purpose of the Study:

• To develop an advanced decoding method for motor imagery EEG signals.
• To enhance the accuracy of brain-computer interfaces (BCI) in neurorehabilitation.
• To address the limitations of traditional deep learning approaches in EEG analysis.

Main Methods:

• Proposed a 3D Convolutional Neural Network (P-3DCNN) to jointly learn spatial-frequency features from EEG signals.
• Utilized Welch's method for frequency band power spectrum calculation and constructed spatial topology matrices.
• Employed cubic interpolation for temporal EEG data and designed a 3DCNN with 1D and 2D convolutional layers, incorporating batch normalization and dropout.

Main Results:

• The P-3DCNN method achieved an average decoding accuracy rate of 86.69% for motor imagery tasks.
• Outperformed various classic machine learning and advanced deep learning techniques in experimental comparisons.
• Demonstrated effective learning of spatial-frequency features from EEG signals.

Conclusions:

• The proposed P-3DCNN method offers a significant advancement in decoding motor imagery EEG.
• This approach enhances the performance of brain-computer interfaces (BCI) for neurorehabilitation applications.
• The findings provide valuable insights for future BCI development and EEG signal processing.